#66 – Michael Cohen on Input Tampering in Advanced RL Agents
Hear This Idea
00:00
How to Give an AI a Model of the World
In this setup, we're like showing an agent a screen with a number on it and asking you to make the number go up. And presumably this is just like likely to be true in practice for our agents doing things in the actual world. In simple environments like chess, you can you can give it the luxury of a known environment where it has the code for for the effects of its actions. If it's acting in the real world or doing most economically useful tasks that I can think of, we won't be able to give it such a model.
Play episode from 22:21
Transcript
